[Feat]: VM and container accounting via /proc/pid rather than cgroups

[shodanshok]

Problem

When using ZFS, cgroups do not correctly account for written data (and, possibly, prefetched read data). Give a look here:

This could be fixed by using normal /proc/pid stats rather than cgroups for gathering these info.

Description

Use classical /proc/pid stats for VM and container statistics.

Importance

really want

Value proposition

1. not depending of cgroups for accounting data is going to help with ZFS

Proposed implementation

No response

[ilyam8]

Hey, @shodanshok. Do you have DefaultIOAccounting=yes in /etc/systemd/system.conf? This fixed the problem for me on my Proxmox servers with ZFS.

After enabling IO accounting:

sudo systemctl daemon-reexec

[ilyam8]

I think this request is wontfix because it will not work for qemu VMs.

[shodanshok]

Hey, @shodanshok. Do you have DefaultIOAccounting=yes in /etc/systemd/system.conf? This fixed the problem for me on my Proxmox servers with ZFS.

@ilyam8 yes, I tried DefaultIOAccounting=yes (and systemctl daemon-reexec + reboot) but nothing changed on my Rocky Linux 9.4. Any idea?

Accounting via /proc/[pid] should work for qemu VMs: while IO is done a dedicated threads, the counters under /proc/[pid]/ioseem to account for them.